Visual Odometry Pose Estimation Based On Point And Line Features Fusion In Dynamic Scenes

With the development of mobile robot technology,the application field has shifted from a more single scene to the direction of unknown complex environment.Unlike traditional robots that need to rely on human-occurring instructions to perform operations,modern mobile robots gradually realize autonomous perception and decision-making,and the issuer of action instructions is the machine itself,so how to autonomously complete various tasks in unknown and complex environments places higher requirements on robot positioning and navigation.In the unknown environment,if mobile robots want to achieve autonomous navigation,the first task to be accomplished is autonomous localization.The traditional wheeled odometry and laser odometry have certain errors in localization.Visual odometry uses visual images for odometry estimation,and the mobile robot uses one or more cameras carried by itself to obtain environmental images at different locations and moments in the environment,and uses the matching of these images to estimate the positional changes of the intelligent body between different moments.Traditional visual odometers are mostly based on point features to build systems.In an environment with a lack of texture,the system usually fails to extract a sufficient number of feature information,resulting in large deviations in the positioning results.The system itself is usually established under strong assumptions of a static environment,that is,it only depends on the movement of the camera itself,and the characteristic information will not move actively.With the continuous expansion of visual odometer application scenarios,applications in dynamic scenarios will inevitably be involved.How to improve the accuracy and robustness of the visual odometer system in dynamic scenarios is of great significance to its future development.The main research contents and research methods of this paper include:First,for the visual odometry system in the under-textured environment,the amount of feature extraction is small,which easily leads to the problem of incorrect pose estimation results.Based on the research of point feature extraction and matching algorithm,line feature information is added,and the pose information of the camera is estimated by fusing point and line features.Then,for the dynamic objects existing in the actual scene,it is proposed to combine the GMS feature matching algorithm with the K-Means clustering algorithm to identify the dynamic area in the system operation scene,and the bad features located in the dynamic object area are analyzed.By filtering,the quantity and quality of characteristic information are retained to the utmost extent,and the occurrence of estimation failures is avoided.In the end,the algorithm in this paper is highly scalable,suitable for monocular,RGB-D and other camera pose estimation,and has been verified many times on the public data set TUM data set.Through experimental comparison with the excellent open source SLAM algorithm,the D-PL algorithm proposed in this paper has higher accuracy and robustness in various experimental indicators.